Share with your CTO
The hyperscalers are betting roughly $750 billion annually on AI infrastructure, and the scaling thesis underneath that number, the idea that bigger models in bigger data centers reliably get smarter, remains unproven. MIT Technology Review’s analysis draws on economists at Sloan and Columbia to argue that a spending retrenchment is not a tail risk but a near-certainty, with the open question being whether it arrives in 2025 or 2028. Unlike the dot-com bust, where infrastructure built during the frenzy outlasted the crash, this cycle has entangled the technology’s own progress with the financial fate of the physical plants being built to run it.
What this means for your business
Whether this story is about your organization depends almost entirely on how much of your AI roadmap sits on top of frontier model capacity rented from the hyperscalers versus workloads you could route to smaller, cheaper, locally run models. Companies that treated “use the biggest model available” as a default architectural decision are the most exposed. Companies that have been quietly evaluating good-enough open-weight models or edge inference, often dismissed as cutting corners, may find themselves in a structurally stronger position if hyperscaler pricing spikes or availability tightens during a retrenchment.
The specific danger the piece names is worth taking seriously: special purpose vehicles (SPVs, off-balance-sheet financing structures similar to those used before the 2008 financial crisis) are reportedly back in the hyperscaler capital stack. If Columbia economist Stijn Van Nieuwerburgh is correct that hyperscaler debt is becoming entangled throughout the broader financial system, a correction in data center valuations wouldn’t stay contained to tech. That’s not a reason to stop building with AI, but it is a reason to pressure-test vendor concentration. A CTO whose production workloads depend on a single hyperscaler’s continued willingness to invest at current rates is carrying counterparty risk that most architecture reviews don’t formally score.
The dot-com parallel is instructive but incomplete. The fiber-optic glut of the late 1990s became genuinely useful infrastructure because bandwidth is bandwidth, a commodity that any application can consume. GPU clusters optimized for training frontier-scale models are far less fungible. If the scaling thesis stalls, those assets don’t quietly become the backbone of the next generation of applications the way dark fiber did. The leading indicator to watch isn’t whether a retrenchment happens but whether benchmark improvements on frontier models continue to justify the marginal cost of the next hardware generation. If that slope flattens before 2027, the architectural bet to revisit isn’t your cloud spend, it’s your dependency on capabilities that only exist at frontier scale.
Concept deep-dive: Scaling thesis
The scaling thesis is the belief that increasing model size, training data, and compute in tandem produces reliably smarter AI, roughly the way adding lanes to a highway increases throughput. It’s the intellectual foundation for every nine-figure data center announcement of the past three years. The business connection is direct: if the thesis holds, today’s infrastructure spending is table stakes. If it hits a wall, the entire capital allocation logic for frontier AI collapses, and competitive advantage shifts toward efficiency and deployment, not raw compute.
Based on reporting from What must happen for AI’s trillion-dollar gamble to pay off, originally published 2026-09-15 06:00:00.
